False Reject Economics: When Your Quality System Is Too Aggressive

A 0.8% false reject rate on a line running 3,000 parts per shift is 24 parts per shift diverted to a reject bin that should have passed. At $12 average part value, that is $288 per shift in unnecessary scrap or rework. Multiply by three shifts and 250 operating days and you have $216,000 per year in waste - all caused by an inspection system that is working exactly as configured, just configured incorrectly.

False rejects are not a sign that your system is strict. They are a sign that your detection threshold is not calibrated to your actual reject specification. The cost is real, it is ongoing, and it is entirely preventable.

What Causes False Rejects

Three root causes account for most false reject problems in deployed vision systems. The first is an overly conservative confidence threshold. During commissioning, there is natural pressure to minimize misses - no one wants a defective part to escape. Thresholds get set low, which catches everything above the boundary but also flags a percentage of borderline-good parts.

The second cause is model brittleness on surface variation. A model trained on parts from one batch of raw material may produce elevated anomaly scores on parts from a different supplier lot with slightly different surface texture. The parts are within specification. The model has not seen that texture distribution. It calls them defective because they are novel, not because they are bad.

The third is lighting instability. Gradual drift in LED intensity, accumulation of dust on the illuminator, or seasonal temperature effects on LED output change the image brightness distribution over time. A model calibrated to the original illumination may produce elevated scores for anomalies that are actually just brightness differences relative to training conditions.

Quantifying the Cost

The direct cost of a false reject depends on what happens to the part. If it goes straight to scrap, the cost is full part value plus disposal. If it goes to a manual review station for confirmation before scrap, the cost is part value plus review labor plus the downstream delay to production flow. If it goes to rework, the cost is rework labor and any yield loss from rework damage.

There is also an indirect cost that is harder to quantify: operator trust erosion. When production operators see the inspection system repeatedly flagging parts that manual review confirms are good, they start overriding the system. An inspection system that operators route around is not an inspection system - it is an annoyance with cameras attached. Recovering from that trust deficit requires demonstrating consistent improvement, which takes time and effort that would not have been necessary if the system had been calibrated correctly at deployment.

The Precision-Recall Tradeoff

Every detection threshold is a point on the precision-recall curve. Moving the threshold up to reduce false rejects (improve precision) will reduce true defect catches if the defect and good-part score distributions overlap in that region. The correct threshold is where the total cost of misses plus the total cost of false rejects is minimized - not where either is individually minimized.

To find that point, you need cost data. What is the cost of a missed defect reaching the customer? For a $12 stamped bracket, the cost is probably $80-$200 in warranty claim handling and potential sorting costs. For a $400 precision component going into a safety-critical assembly, it could be $5,000-$50,000 in downstream liability. What is the cost of a false reject? Part value plus disposition cost.

If your miss cost is 20x your false reject cost, your threshold should be set to allow up to 20 false rejects for every miss prevented. That is the economically rational calibration. Systems calibrated purely on detection rate without this cost weighting are almost always set too conservatively.

Practical Calibration Process

Threshold calibration requires a validation dataset with known outcomes - a set of parts that have been confirmed as good through manual inspection or downstream functional testing, mixed with confirmed defective parts. Run the full validation set through the system at multiple threshold levels and record true accept rate, false reject rate, true reject rate, and false accept rate at each threshold.

Plotting these as a cost curve using your actual cost values will show you the optimal operating point. For most industrial inspection applications, the optimal threshold is not the default setting established during initial calibration. Systems should be recalibrated every six months or when process changes alter part appearance.

Model Retraining vs Threshold Adjustment

Threshold adjustment is fast and reversible. If your false reject rate is elevated but your miss rate is acceptable, try raising the threshold first before committing to model retraining. If raising the threshold reduces false rejects without meaningfully increasing misses, the model is fine - you were just running it too conservatively.

If raising the threshold to reduce false rejects also increases misses proportionally, the model is not separating defect and good-part score distributions cleanly in the problem region. That requires retraining with better representation of the borderline cases - adding more labeled samples of the surface variation that is causing false rejects, so the model can learn to distinguish it from actual defects.

Retraining is the right solution when the model is structurally wrong. Threshold adjustment is the right solution when the model is right but operating parameters are off. Running retraining when a threshold adjustment would suffice wastes engineering time and introduces model change risk unnecessarily.

Monitoring False Reject Rate Continuously

False reject rate should be a tracked metric on your quality dashboard alongside detection rate and escape rate. A system that was running at 0.3% false rejects and is now running at 1.1% has changed. Something upstream changed - illumination, material batch, part geometry from tooling wear - and the system is responding to it. Catching that drift early, before it becomes an operator trust problem or a significant cost accumulation, requires active monitoring rather than periodic audits.